kermitt2 / grobid

A machine learning software for extracting information from scholarly documents
https://grobid.readthedocs.io
Apache License 2.0
3.59k stars 459 forks source link

Empty refs #1175

Open lfoppiano opened 1 month ago

lfoppiano commented 1 month ago

I've stumbled upon a strange bug, here the input:

image

and here the output

<s>Table 
                        <ref type="table" target="#tab_0">I</ref> gives the coefficients a and b, and the standard deviations of the linear fitted lines at T C for magnetostriction DL/L vs M 2 and DL/L vs M 4 , as shown in Figs. 
                        <ref type="figure">5</ref> and
                        <ref type="figure"/>  respectively.
                    </s>

This occurs both with or without sentence segmentation:

<p>, where C h is the magnetovolume (magnetoelastic) constant, M 0 2 (T) is the spontaneous magnetization at the measured temperature, and M is the magnetization under the magnetic field. 
                    <ref type="bibr" target="#b5">6</ref> Kittel said that the volume strain DV/V is proportional to e xx þ e yy þ e zz , where e xx , e yy , and e zz indicate the strains for the x, y, and z axes, respectively. 
                    <ref type="bibr" target="#b14">15</ref> We used a polycrystalline sample; therefore, the DV/V is equal to three times DL/L. Accordingly, the volume magnetostriction DV/V is proportional to the magnetostriction DL/L. As a result, discussions about the spin fluctuation theory suggested by Takahashi can be applied to the results of the magnetostriction DL/L in this experimental study. Takahashi suggested that the magnetovolume constant C h decreases in proportion to jT À T C j. It is suggested that as C h decreases to zero toward the Curie temperature, r 4 (M 4 )-linear behavior appears, as shown in Eq. (1). Table 
                    <ref type="table" target="#tab_0">I</ref> gives the coefficients a and b, and the standard deviations of the linear fitted lines at T C for magnetostriction DL/L vs M 2 and DL/L vs M 4 , as shown in Figs. 
                    <ref type="figure">5</ref> and
                    <ref type="figure"/>  respectively. At T C , the fitted linear line is given by the equation of DL/L ¼ a þ bM d for d ¼ 2 or 4. The errors of the coefficient b were within 6 2% for both values of d. The fractions of the coefficient a and the magnetostriction at 5 T, DL/L ¼ À50 Â 10 À6 , were 56% and 0.68% for d ¼ 2 and 4, respectively. This analysis indicates that the y-intercept of DL/L vs M 4 , as shown in Fig. 
                    <ref type="figure">6</ref>, is almost zero and the magnetostriction can be represented by the equation DL/L ¼ bM 4 at the temperature T C .
                </p>

This case occurs within the table caption, may be more frequent, not sure about that.

lfoppiano commented 1 month ago

This seems to happen only on the Docker-ran version on huggingface. Natively on M2 works fine.