Add Q&A and Book, fix task running bugs (#50)

rag/app/__init__.py

@@ -4,14 +4,8 @@ from nltk import word_tokenize
 
 from rag.nlp import stemmer, huqie
 
-
-def callback__(progress, msg, func):
-    if not func :return
-    func(progress, msg)
-
-
 BULLET_PATTERN = [[
-        r"第[零一二三四五六七八九十百]+编",
+        r"第[零一二三四五六七八九十百]+(编|部分)",
         r"第[零一二三四五六七八九十百]+章",
         r"第[零一二三四五六七八九十百]+节",
         r"第[零一二三四五六七八九十百]+条",
@@ -22,6 +16,8 @@ BULLET_PATTERN = [[
         r"[0-9]{,2}\.[0-9]{,2}\.[0-9]{,2}",
         r"[0-9]{,2}\.[0-9]{,2}\.[0-9]{,2}\.[0-9]{,2}",
     ], [
+        r"第[零一二三四五六七八九十百]+章",
+        r"第[零一二三四五六七八九十百]+节",
         r"[零一二三四五六七八九十百]+[ 、]",
         r"[\(（][零一二三四五六七八九十百]+[\)）]",
         r"[\(（][0-9]{,2}[\)）]",
@@ -54,7 +50,7 @@ def bullets_category(sections):
 def is_english(texts):
     eng = 0
     for t in texts:
-        if re.match(r"[a-zA-Z]", t.strip()):
+        if re.match(r"[a-zA-Z]{2,}", t.strip()):
             eng += 1
     if eng / len(texts) > 0.8:
         return True
@@ -70,3 +66,26 @@ def tokenize(d, t, eng):
         d["content_sm_ltks"] = huqie.qieqie(d["content_ltks"])
 
 
+def remove_contents_table(sections, eng=False):
+    i = 0
+    while i < len(sections):
+        def get(i):
+            nonlocal sections
+            return (sections[i] if type(sections[i]) == type("") else sections[i][0]).strip()
+        if not re.match(r"(contents|目录|目次|table of contents|致谢|acknowledge)$", re.sub(r"( | |\u3000)+", "", get(i).split("@@")[0], re.IGNORECASE)):
+            i += 1
+            continue
+        sections.pop(i)
+        if i >= len(sections): break
+        prefix = get(i)[:3] if not eng else " ".join(get(i).split(" ")[:2])
+        while not prefix:
+            sections.pop(i)
+            if i >= len(sections): break
+            prefix = get(i)[:3] if not eng else " ".join(get(i).split(" ")[:2])
+        sections.pop(i)
+        if i >= len(sections) or not prefix: break
+        for j in range(i, min(i+128, len(sections))):
+            if not re.match(prefix, get(j)):
+                continue
+            for _ in range(i, j):sections.pop(i)
+            break

rag/app/book.py

@@ -0,0 +1,156 @@
+import copy
+import random
+import re
+from io import BytesIO
+from docx import Document
+import numpy as np
+from rag.app import bullets_category, BULLET_PATTERN, is_english, tokenize, remove_contents_table
+from rag.nlp import huqie
+from rag.parser.docx_parser import HuDocxParser
+from rag.parser.pdf_parser import HuParser
+
+
+class Pdf(HuParser):
+    def __call__(self, filename, binary=None, from_page=0,
+                 to_page=100000, zoomin=3, callback=None):
+        self.__images__(
+            filename if not binary else binary,
+            zoomin,
+            from_page,
+            to_page)
+        callback(0.1, "OCR finished")
+
+        from timeit import default_timer as timer
+        start = timer()
+        self._layouts_paddle(zoomin)
+        callback(0.47, "Layout analysis finished")
+        print("paddle layouts:", timer() - start)
+        self._table_transformer_job(zoomin)
+        callback(0.68, "Table analysis finished")
+        self._text_merge()
+        column_width = np.median([b["x1"] - b["x0"] for b in self.boxes])
+        self._concat_downward(concat_between_pages=False)
+        self._filter_forpages()
+        self._merge_with_same_bullet()
+        callback(0.75, "Text merging finished.")
+        tbls = self._extract_table_figure(True, zoomin, False)
+
+        callback(0.8, "Text extraction finished")
+
+        return [(b["text"] + self._line_tag(b, zoomin), b.get("layoutno","")) for b in self.boxes]
+
+
+def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None):
+    doc = {
+        "docnm_kwd": filename,
+        "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))
+    }
+    doc["title_sm_tks"] = huqie.qieqie(doc["title_tks"])
+    pdf_parser = None
+    sections,tbls = [], []
+    if re.search(r"\.docx?$", filename, re.IGNORECASE):
+        callback(0.1, "Start to parse.")
+        doc_parser = HuDocxParser()
+        # TODO: table of contents need to be removed
+        sections, tbls = doc_parser(binary if binary else filename)
+        remove_contents_table(sections, eng = is_english(random.choices([t for t,_ in sections], k=200)))
+        callback(0.8, "Finish parsing.")
+    elif re.search(r"\.pdf$", filename, re.IGNORECASE):
+        pdf_parser = Pdf()
+        sections,tbls = pdf_parser(filename if not binary else binary,
+                         from_page=from_page, to_page=to_page, callback=callback)
+    elif re.search(r"\.txt$", filename, re.IGNORECASE):
+        callback(0.1, "Start to parse.")
+        txt = ""
+        if binary:txt = binary.decode("utf-8")
+        else:
+            with open(filename, "r") as f:
+                while True:
+                    l = f.readline()
+                    if not l:break
+                    txt += l
+            sections = txt.split("\n")
+        sections = [(l,"") for l in sections if l]
+        remove_contents_table(sections, eng = is_english(random.choices([t for t,_ in sections], k=200)))
+        callback(0.8, "Finish parsing.")
+    else: raise NotImplementedError("file type not supported yet(docx, pdf, txt supported)")
+
+    bull = bullets_category([b["text"] for b in random.choices([t for t,_ in sections], k=100)])
+    projs = [len(BULLET_PATTERN[bull]) + 1] * len(sections)
+    levels = [[]] * len(BULLET_PATTERN[bull]) + 2
+    for i, (txt, layout) in enumerate(sections):
+        for j, p in enumerate(BULLET_PATTERN[bull]):
+            if re.match(p, txt.strip()):
+                projs[i] = j
+                levels[j].append(i)
+                break
+        else:
+            if re.search(r"(title|head)", layout):
+                projs[i] = BULLET_PATTERN[bull]
+                levels[BULLET_PATTERN[bull]].append(i)
+            else:
+                levels[BULLET_PATTERN[bull] + 1].append(i)
+    sections = [t for t,_ in sections]
+
+    def binary_search(arr, target):
+        if target > arr[-1]: return len(arr) - 1
+        if target > arr[0]: return -1
+        s, e = 0, len(arr)
+        while e - s > 1:
+            i = (e + s) // 2
+            if target > arr[i]:
+                s = i
+                continue
+            elif target < arr[i]:
+                e = i
+                continue
+            else:
+                assert False
+        return s
+
+    cks = []
+    readed = [False] * len(sections)
+    levels = levels[::-1]
+    for i, arr in enumerate(levels):
+        for j in arr:
+            if readed[j]: continue
+            readed[j] = True
+            cks.append([j])
+            if i + 1 == len(levels) - 1: continue
+            for ii in range(i + 1, len(levels)):
+                jj = binary_search(levels[ii], j)
+                if jj < 0: break
+                if jj > cks[-1][-1]: cks[-1].pop(-1)
+                cks[-1].append(levels[ii][jj])
+
+    # is it English
+    eng = is_english(random.choices(sections, k=218))
+
+    res = []
+    # add tables
+    for img, rows in tbls:
+        bs = 10
+        de = ";" if eng else "；"
+        for i in range(0, len(rows), bs):
+            d = copy.deepcopy(doc)
+            r = de.join(rows[i:i + bs])
+            r = re.sub(r"\t——(来自| in ).*”%s" % de, "", r)
+            tokenize(d, r, eng)
+            d["image"] = img
+            res.append(d)
+    # wrap up to es documents
+    for ck in cks:
+        print("\n-".join(ck[::-1]))
+        ck = "\n".join(ck[::-1])
+        d = copy.deepcopy(doc)
+        if pdf_parser:
+            d["image"] = pdf_parser.crop(ck)
+            ck = pdf_parser.remove_tag(ck)
+        tokenize(d, ck, eng)
+        res.append(d)
+    return res
+
+
+if __name__ == "__main__":
+    import sys
+    chunk(sys.argv[1])

rag/app/laws.py

@@ -3,7 +3,7 @@ import re
 from io import BytesIO
 from docx import Document
 import numpy as np
-from rag.app import callback__, bullets_category, BULLET_PATTERN, is_english, tokenize
+from rag.app import bullets_category, BULLET_PATTERN, is_english, tokenize
 from rag.nlp import huqie
 from rag.parser.docx_parser import HuDocxParser
 from rag.parser.pdf_parser import HuParser
@@ -32,12 +32,12 @@ class Pdf(HuParser):
             zoomin,
             from_page,
             to_page)
-        callback__(0.1, "OCR finished", callback)
+        callback(0.1, "OCR finished")
 
         from timeit import default_timer as timer
         start = timer()
         self._layouts_paddle(zoomin)
-        callback__(0.77, "Layout analysis finished", callback)
+        callback(0.77, "Layout analysis finished")
         print("paddle layouts:", timer()-start)
         bxs = self.sort_Y_firstly(self.boxes, np.median(self.mean_height) / 3)
         # is it English
@@ -75,7 +75,7 @@ class Pdf(HuParser):
             b["x1"] = max(b["x1"], b_["x1"])
             bxs.pop(i + 1)
 
-        callback__(0.8, "Text extraction finished", callback)
+        callback(0.8, "Text extraction finished")
 
         return [b["text"] + self._line_tag(b, zoomin) for b in bxs]
 
@@ -89,17 +89,17 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None):
     pdf_parser = None
     sections = []
     if re.search(r"\.docx?$", filename, re.IGNORECASE):
-        callback__(0.1, "Start to parse.", callback)
+        callback(0.1, "Start to parse.")
         for txt in Docx()(filename, binary):
             sections.append(txt)
-        callback__(0.8, "Finish parsing.", callback)
+        callback(0.8, "Finish parsing.")
     elif re.search(r"\.pdf$", filename, re.IGNORECASE):
         pdf_parser = Pdf()
         for txt in pdf_parser(filename if not binary else binary,
                          from_page=from_page, to_page=to_page, callback=callback):
             sections.append(txt)
     elif re.search(r"\.txt$", filename, re.IGNORECASE):
-        callback__(0.1, "Start to parse.", callback)
+        callback(0.1, "Start to parse.")
         txt = ""
         if binary:txt = binary.decode("utf-8")
         else:
@@ -110,7 +110,7 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None):
                     txt += l
             sections = txt.split("\n")
         sections = [l for l in sections if l]
-        callback__(0.8, "Finish parsing.", callback)
+        callback(0.8, "Finish parsing.")
     else: raise NotImplementedError("file type not supported yet(docx, pdf, txt supported)")
 
     # is it English
@@ -118,7 +118,7 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None):
     # Remove 'Contents' part
     i = 0
     while i < len(sections):
-        if not re.match(r"(Contents|目录|目次)$", re.sub(r"( | |\u3000)+", "", sections[i].split("@@")[0])):
+        if not re.match(r"(contents|目录|目次|table of contents)$", re.sub(r"( | |\u3000)+", "", sections[i].split("@@")[0], re.IGNORECASE)):
             i += 1
             continue
         sections.pop(i)
@@ -133,7 +133,7 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None):
         for j in range(i, min(i+128, len(sections))):
             if not re.match(prefix, sections[j]):
                 continue
-            for k in range(i, j):sections.pop(i)
+            for _ in range(i, j):sections.pop(i)
             break
 
     bull = bullets_category(sections)

rag/app/manual.py

@@ -1,6 +1,6 @@
 import copy
 import re
-from rag.app import callback__, tokenize
+from rag.app import tokenize
 from rag.nlp import huqie
 from rag.parser.pdf_parser import HuParser
 from rag.utils import num_tokens_from_string
@@ -14,19 +14,19 @@ class Pdf(HuParser):
             zoomin,
             from_page,
             to_page)
-        callback__(0.2, "OCR finished.", callback)
+        callback(0.2, "OCR finished.")
 
         from timeit import default_timer as timer
         start = timer()
         self._layouts_paddle(zoomin)
-        callback__(0.5, "Layout analysis finished.", callback)
+        callback(0.5, "Layout analysis finished.")
         print("paddle layouts:", timer() - start)
         self._table_transformer_job(zoomin)
-        callback__(0.7, "Table analysis finished.", callback)
+        callback(0.7, "Table analysis finished.")
         self._text_merge()
         self._concat_downward(concat_between_pages=False)
         self._filter_forpages()
-        callback__(0.77, "Text merging finished", callback)
+        callback(0.77, "Text merging finished")
         tbls = self._extract_table_figure(True, zoomin, False)
 
         # clean mess
@@ -34,20 +34,8 @@ class Pdf(HuParser):
             b["text"] = re.sub(r"([\t 　]|\u3000){2,}", " ", b["text"].strip())
 
         # merge chunks with the same bullets
-        i = 0
-        while i + 1 < len(self.boxes):
-            b = self.boxes[i]
-            b_ = self.boxes[i + 1]
-            if b["text"].strip()[0] != b_["text"].strip()[0] \
-                    or b["page_number"]!=b_["page_number"] \
-                    or b["top"] > b_["bottom"]:
-                i += 1
-                continue
-            b_["text"] = b["text"] + "\n" + b_["text"]
-            b_["x0"] = min(b["x0"], b_["x0"])
-            b_["x1"] = max(b["x1"], b_["x1"])
-            b_["top"] = b["top"]
-            self.boxes.pop(i)
+        self._merge_with_same_bullet()
+
         # merge title with decent chunk
         i = 0
         while i + 1 < len(self.boxes):
@@ -62,7 +50,7 @@ class Pdf(HuParser):
             b_["top"] = b["top"]
             self.boxes.pop(i)
 
-        callback__(0.8, "Parsing finished", callback)
+        callback(0.8, "Parsing finished")
         for b in self.boxes: print(b["text"], b.get("layoutno"))
 
         print(tbls)

rag/app/paper.py

@@ -1,11 +1,9 @@
 import copy
 import re
 from collections import Counter
-from rag.app import callback__, bullets_category, BULLET_PATTERN, is_english, tokenize
-from rag.nlp import huqie, stemmer
-from rag.parser.docx_parser import HuDocxParser
+from rag.app import tokenize
+from rag.nlp import huqie
 from rag.parser.pdf_parser import HuParser
-from nltk.tokenize import word_tokenize
 import numpy as np
 from rag.utils import num_tokens_from_string
 
@@ -18,20 +16,20 @@ class Pdf(HuParser):
             zoomin,
             from_page,
             to_page)
-        callback__(0.2, "OCR finished.", callback)
+        callback(0.2, "OCR finished.")
 
         from timeit import default_timer as timer
         start = timer()
         self._layouts_paddle(zoomin)
-        callback__(0.47, "Layout analysis finished", callback)
+        callback(0.47, "Layout analysis finished")
         print("paddle layouts:", timer() - start)
         self._table_transformer_job(zoomin)
-        callback__(0.68, "Table analysis finished", callback)
+        callback(0.68, "Table analysis finished")
         self._text_merge()
         column_width = np.median([b["x1"] - b["x0"] for b in self.boxes])
         self._concat_downward(concat_between_pages=False)
         self._filter_forpages()
-        callback__(0.75, "Text merging finished.", callback)
+        callback(0.75, "Text merging finished.")
         tbls = self._extract_table_figure(True, zoomin, False)
 
         # clean mess
@@ -101,7 +99,7 @@ class Pdf(HuParser):
                 break
         if not abstr: i = 0
 
-        callback__(0.8, "Page {}~{}: Text merging finished".format(from_page, min(to_page, self.total_page)), callback)
+        callback(0.8, "Page {}~{}: Text merging finished".format(from_page, min(to_page, self.total_page)))
         for b in self.boxes: print(b["text"], b.get("layoutno"))
         print(tbls)
 

rag/app/presentation.py

@@ -3,7 +3,7 @@ import re
 from io import BytesIO
 from pptx import Presentation
 
-from rag.app import callback__, tokenize, is_english
+from rag.app import tokenize, is_english
 from rag.nlp import huqie
 from rag.parser.pdf_parser import HuParser
 
@@ -43,7 +43,7 @@ class Ppt(object):
                 if txt: texts.append(txt)
             txts.append("\n".join(texts))
 
-        callback__(0.5, "Text extraction finished.", callback)
+        callback(0.5, "Text extraction finished.")
         import aspose.slides as slides
         import aspose.pydrawing as drawing
         imgs = []
@@ -53,7 +53,7 @@ class Ppt(object):
                 slide.get_thumbnail(0.5, 0.5).save(buffered, drawing.imaging.ImageFormat.jpeg)
                 imgs.append(buffered.getvalue())
         assert len(imgs) == len(txts), "Slides text and image do not match: {} vs. {}".format(len(imgs), len(txts))
-        callback__(0.9, "Image extraction finished", callback)
+        callback(0.9, "Image extraction finished")
         self.is_english = is_english(txts)
         return [(txts[i], imgs[i]) for i in range(len(txts))]
 
@@ -70,7 +70,7 @@ class Pdf(HuParser):
 
     def __call__(self, filename, binary=None, from_page=0, to_page=100000, zoomin=3, callback=None):
         self.__images__(filename if not binary else binary, zoomin, from_page, to_page)
-        callback__(0.8, "Page {}~{}: OCR finished".format(from_page, min(to_page, self.total_page)), callback)
+        callback(0.8, "Page {}~{}: OCR finished".format(from_page, min(to_page, self.total_page)))
         assert len(self.boxes) == len(self.page_images), "{} vs. {}".format(len(self.boxes), len(self.page_images))
         res = []
         #################### More precisely ###################
@@ -89,7 +89,7 @@ class Pdf(HuParser):
         for i in range(len(self.boxes)):
             lines = "\n".join([b["text"] for b in self.boxes[i] if not self.__garbage(b["text"])])
             res.append((lines, self.page_images[i]))
-        callback__(0.9, "Page {}~{}: Parsing finished".format(from_page, min(to_page, self.total_page)), callback)
+        callback(0.9, "Page {}~{}: Parsing finished".format(from_page, min(to_page, self.total_page)))
         return res
 
 

rag/app/qa.py

@@ -0,0 +1,104 @@
+import random
+import re
+from io import BytesIO
+from nltk import word_tokenize
+from openpyxl import load_workbook
+from rag.app import is_english
+from rag.nlp import huqie, stemmer
+
+
+class Excel(object):
+    def __call__(self, fnm, binary=None, callback=None):
+        if not binary:
+            wb = load_workbook(fnm)
+        else:
+            wb = load_workbook(BytesIO(binary))
+        total = 0
+        for sheetname in wb.sheetnames:
+            total += len(list(wb[sheetname].rows))
+
+        res, fails = [], []
+        for sheetname in wb.sheetnames:
+            ws = wb[sheetname]
+            rows = list(ws.rows)
+            for i, r in enumerate(rows):
+                q, a = "", ""
+                for cell in r:
+                    if not cell.value: continue
+                    if not q: q = str(cell.value)
+                    elif not a: a = str(cell.value)
+                    else: break
+                if q and a: res.append((q, a))
+                else: fails.append(str(i+1))
+                if len(res) % 999 == 0:
+                    callback(len(res)*0.6/total, ("Extract Q&A: {}".format(len(res)) + (f"{len(fails)} failure, line: %s..."%(",".join(fails[:3])) if fails else "")))
+
+        callback(0.6, ("Extract Q&A: {}".format(len(res)) + (
+            f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
+        self.is_english = is_english([rmPrefix(q) for q, _ in random.choices(res, k=30) if len(q)>1])
+        return res
+
+
+def rmPrefix(txt):
+    return re.sub(r"^(问题|答案|回答|user|assistant|Q|A|Question|Answer|问|答)[\t:： ]+", "", txt.strip(), flags=re.IGNORECASE)
+
+
+def beAdoc(d, q, a, eng):
+    qprefix = "Question: " if eng else "问题："
+    aprefix = "Answer: " if eng else "回答："
+    d["content_with_weight"] = "\t".join([qprefix+rmPrefix(q), aprefix+rmPrefix(a)])
+    if eng:
+        d["content_ltks"] = " ".join([stemmer.stem(w) for w in word_tokenize(q)])
+    else:
+        d["content_ltks"] = huqie.qie(q)
+        d["content_sm_ltks"] = huqie.qieqie(d["content_ltks"])
+    return d
+
+
+def chunk(filename, binary=None,  from_page=0, to_page=100000, callback=None):
+
+    res = []
+    if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
+        callback(0.1, "Start to parse.")
+        excel_parser = Excel()
+        for q,a in excel_parser(filename, binary, callback):
+            res.append(beAdoc({}, q, a, excel_parser.is_english))
+        return res
+    elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
+        callback(0.1, "Start to parse.")
+        txt = ""
+        if binary:
+            txt = binary.decode("utf-8")
+        else:
+            with open(filename, "r") as f:
+                while True:
+                    l = f.readline()
+                    if not l: break
+                    txt += l
+        lines = txt.split("\n")
+        eng = is_english([rmPrefix(l) for l in lines[:100]])
+        fails = []
+        for i, line in enumerate(lines):
+            arr = [l for l in line.split("\t") if len(l) > 1]
+            if len(arr) != 2:
+                fails.append(str(i))
+                continue
+            res.append(beAdoc({}, arr[0], arr[1], eng))
+            if len(res) % 999 == 0:
+                callback(len(res) * 0.6 / len(lines), ("Extract Q&A: {}".format(len(res)) + (
+                    f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
+
+        callback(0.6, ("Extract Q&A: {}".format(len(res)) + (
+            f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
+
+        return res
+
+    raise NotImplementedError("file type not supported yet(pptx, pdf supported)")
+
+
+if __name__== "__main__":
+    import sys
+    def kk(rat, ss):
+        pass
+    print(chunk(sys.argv[1], callback=kk))
+